Research Of Fall Detection Based On Pose Estimation

With the advent of an aging era,more and more empty nesters will be added to the society.How to protect the health and safety of empty nesters is a considered question.According to the survey,the fall of empty nesters is an important factor threatening their health and safety.If the elderly can get help in time after the fall,it will greatly reduce some following diseases caused by the fall.Based on the above reasons,how to design a real-time and robust fall detection method has always been a hot topic.At present,the research on the fall detection method can be roughly divided into three directions: the wearable fall detection method,which collects the sports information of the elderly by the elderly wearing information collecting equipment,such as accelerometers and gyroscopes.Detection and judgment,the advantage of this method is that there is no limit to activities of the elderly,whose disadvantage is that the device needs to be worn at any time.If the elderly forgets to wear the device,the fall will not be detected.The fall detection method based on peripheral sensor mainly collects the activity information of human body by wearing pressure sensor,infrared sensor and other equipment.The advantage is that the elderly is not required to wear the equipment at any time,which improves the comfort of the elderly.The shortcomings are sensitive to noise and are prone to incorrect detection.The fall detection method based on video surveillance mainly collects human body information through the camera.The advantage is that the device is easy to acquire and insensitive to noise and has a high detection accuracy.This paper mainly studies the fall detection method in the video surveillance scene.The work of this paper is as follows:1.Research and analysis of traditional fall detection methods.Firstly,the mixed Gaussian background model is used to extract the foreground,and the extracted foreground is morphologically processed.After obtaining the foreground of the human body,it is necessary to extract the fall feature of the foreground.In this paper,Hu invariant moment,external moment aspect ratio,attitude change rate and motion rate are used as the features of fall detection.Finally,the SVM classifier is trained using the extracted features to complete the recognition of the fall activity.2.The traditional fall detection method is to calculate the fall detection by calculating the shape and motion features of the human body in the continuous image sequence.It relies on the relationship between the human body context in the continuous image sequence.When the fall detection is performed on the multi-channel video,the computational overhead of the system is relatively serious.In real world,the fall detection system is generally 32 or 64 channels of video.If the typical fall detection method is used,system overload calculation is likely to occur.Aiming at this situation,this paper proposes a fall detection method based on pose estimation,which can judge the fall detection of static pictures,which greatly reduces the computational overhead.The fall detection method in this paper firstly estimates the posture of the human body in the image,and acquires the coordinates of the 18 bone joint points of the human body;for different postures of the human body,the relative positional relationship of the human bone joint points is also different.Therefore,the key points of the human skeleton can be used as the feature vector of the fall judgment,and the fall detection detection is performed using the SVM classifier.3.In practical applications,it is found that when a person is comfortable lying on the seat,the posture information of the human body and the posture information at the time of the fall are similar,and the erroneous detection is apt to occur.In response to this special scenario,this paper uses the YOLOv3 convolutional neural network framework to train a classifier,which is used in conjunction with the above-mentioned pose estimation fall detection method for fall detection.The experimental results show that this can be solved well after adding the YOLOv3 network model and improve the accuracy of the detection.